Legally it is quite clear. Taking a description of a closed source program and writing a new one is ok in most cases (unless that description is API docs - see Cisco vs Arista). Taking a look at closed source software and then implementing your own version is poison as far as OSS goes. OP implemented the first version, so that’s already a problem. They may get away is they describe what the program does to someone else and let them implement it, but OP would not be able to touch the source code

My point is Kubernetes is a hack (a useful hack!) to synchronize multiple separate, different systems in certain ways. It cannot provide anything close to something like a single system image and it can’t bridge the discrete model of computation that Unix assumes.

Kubernetes is not intended to provide anything like a single system image. It’s a workload orchestration system, not an operating system. Given a compatible interface (a runtime) Kubernetes can in theory distribute workloads to any OS.

All these features require a lot of code and complexity to maintain (latest info I can find is almost 2 million as of 2018). Ideally, Kubernetes is capable of what you said, in the same way that ideally programs can’t violate Unix filesystem DAC or other user permissions but in practice every line of code is another opportunity for something to go wrong…

Just because something has more security features doesn’t mean it’s actually secure. Or that it’s maintainable without a company with thousands of engineers and tons of money maintaining for you. Keeping you in a dependent relationship.

I’m not going to argue that Kubernetes is not complex. But as I stated previously Kubernetes as a bespoke ecosystem is less complex than configuring the same features with decoupled systems. The requirements for an orchestrator and the challenges (technical, security, human, etc) to manage said orchestrator are higher. All else being equal, Kubernetes has implemented this in a very lean way, delegating networking, storage, and runtime to pluggable providers on the left, and delegating non-basic workload aspects to operators on the right. It’s this extensibility that makes it both popular with operators and makes it appear daunting to a layperson. And going back to security, is has provably shown to have a reduced attack surface when managed by a competent operator.

So? I don’t expect many of these ideas will be adopted in the mainstream under the monopoly-capitalist market system. It’s way more profitable to keep selling support to manage sprawling and complex systems that require armies of software engineers to upkeep. I think if state investment or public research in general becomes relevant again maybe these ideas will be investigated and adopted for their technical merit.

So you’re… what, dismissing HTTP because it has been adopted by capitalist market systems? Are you going to dismiss the Fediverse for using HTTP? What about widely adopted protocols? DNS, BGP, IPv4/6, etc?

How about we bring this part of the discussion back to the roots? You said that HTTP and REST as communication protocols seemed strange to you because Unix has other primitives. I pointed out that those primitives do not address many modern client-server communication requirements. You did not refute that, but you said, and I paraphrase “9P did it better”. I refrain from commenting on that because there’s no comparative implementation of complex Internet-based systems in 9P. I did state though that even if 9P is superior, as you claim, it did not win out in the end. There’s plenty of precedents for this: Betamax-VHS, git-mercurial, etc.

“Highly available” is carrying a lot of weight there lol. If we can move some of these qualities into a filesystem layer (which is a userspace application on some systems) and get these benefits for free for all data, why shouldn’t we? The filesystem layer and application layer are not 2 fundamentally separate unrelated parts of a whole.

(My emphasis) It’s not free though. There’s an overhead for doing this, and you end up doing things in-filesystem that have no business being there.

It’s not a flawed anecdote or a preconception. They had their own personal experience with a cloud tool and didn’t like it.

*Ahem*:

“Nobody really uses Kubernetes for day-to-day work, and it shows.”

That is not an experience, it’s a provably wrong statement.

The assumption among certain computer touchers is that you can’t use Kubernetes or “cloud” tools and not come away loving them. So if someone doesn’t like them they must not really understand them!

That’s a very weird assumption, and it’s the first time I’ve heard it. Can you provide a source? Because in my experience the opposite is the case - there’s no community more critical of Kubernetes’ flaws than their developers/users themselves.

They probably could’ve said it nicer. It’s still no excuse to dismiss criticism because you didn’t like the tone.

I dismissed the criticism because it makes an appeal to pathos, not to logos. Like I said, there’s plenty of valid technical criticisms of Kubernetes, and even an argument on the basis of ethics (like you’re making) is more engaging.

I think Kubernetes has its uses, for now. But it’s still a fundamentally limited and harmful (because of its monopolistic maintainers/creators) way to do a kind of distributed computing. I don’t think anyone is coming for you to take your Kubernetes though…

No my Kubernetes. I use it because it’s academically interesting, and because it does the tasks it is meant to do better than most alternatives. But if CNCF were to implode today and Kubernetes became no longer practical to use then I would just pivot to another system.

I’m not going to argue whether it’s a harmful way of doing distributed computing based on their maintainers/pedrigee. That’s a longer philosophical discussion than I suspect neither you or I have time for.

Probably need to keep in mind incidental versus essential complexity here.

Go on…

Because this is how much of what we use already is implemented. Significant effort goes in to portability, interoperability and balancing compromises. When I’m doing software development e.g. writing HTTP APIs (of which I apparently know nothing about ;) ) - I feel like I’ve got a responsibility to carefully balance what I expose as some user-configurable thing versus something managed internally by the application. Sometimes, thankfully, the application doesn’t even have to think about it al all - like what TCP flags to set when I dial some service.

In the case of vmalert, the binary makes no assumptions as to default behaviour because it was not meant to be run standalone. It comes as part of a container with specific environment variables, which in turn is packaged as a Helm chart which has sane configurations. Taking the vmalert binary by itself is like taking a kerberos server binary without its libraries and config files in /etc files and complaining that it’s not working.

You bring up containers which is a great example of some cool features provided by the Linux kernel to solve interesting problems. If you’re interested, have a look at FreeBSD’s Jails, Plan 9 and LXC. Compare the interface to all these systems, both at the library level and userspace, and compare the applications developed using those systems. How easy is it to get going? How much do I need to keep in my head when using these features? Docker, Kubernetes, and the rest all have made different tradeoffs and compromises.

I am very well versed in jails, chroot, openvz, LXC, etc. OCI containers are in a different class - don’t think of them as an OS-like environment, think of them as a self-contained, packaged service. Docker is then one example of a runtime runtime on which those services run, and Kubernetes is an orchestrator that managed containers in runtimes. And yes, there are some tradeoffs and compromises, but those are well within the bounds of the Pareto principle - remove the 10% long tail of features on the host, reduce user-facing complexity by 90%.

Another one I think about is SQLite. Some seriously clever smarts. Huge numbers of people don’t know anything about for-loops, C, or B-Trees but can read & write SQL. That’s technology at its best.

Are you arguing that Kubernetes doesn’t do that for you? Because with Kubernetes I can say “run the service in this container with these settings and so many replicas”, attach some conditions like “stop sending traffic to any one container that takes longer than N seconds to respond” and “restart the container if a certain command returns an error”, and just let it run. I can do a rolling upgrade of the nodes and Kubernetes will reschedule the containers on any other available node, it can load balance traffic, I can update the spec of a deployment and Kubernetes will do a zero-downtime upgrade for me. Try implementing the same on a Unix system. You’d need a way to push configs (Ansible, Puppet, etc?). You need load balancing and leader election (Keepalived?). You need error detection. You need DNS. You need to run the services. You need to ensure there’s no library conflict. There’s a LOT of complexity that a Kubernetes user does not need to worry about any more. Tell me that’s not serious smarts and technology at its best.

What I’m struggling with are thoughts of significant vested commercial interest in exposing this kind of detail, fuelling multi-billion dollar service industries. Feelings of being an outsider despite understanding how it all fits together.

You seem to be conflating Kubernetes and cloud services. Being a cloud native technology does not mean it has to run on a managed cloud service. It just means that it has certain expectations as to how workloads run on it, and if those expectations are met then it makes certain promises about how it will behave.

Have you ever written this kind of software before?

I have contributed to several similar open source projects, yes. What about it?

It sounds like you are comfortable with the status quo of this part of the software industry, and I’m truly jealous!

I am comfortable with my knowledge of this part of the software industry. There is no status quo - there’s currently an equilibrium, yes, but it is a tenuous one. I know the tools I use today will likely not be the same tools I will be using a decade from now. But I also know that the concepts and architectures I learn from managing these tools will still be applicable then, and I can stay agile enough to adapt and become comfortable in a new ecosystem. I would urge you to consider the same approach for yourself.

A lot of simple parts build up in predictable ways to accomplish big things. The complexity is spread out and minimized.

This has always felt untrue to me. The command line has always been simple parts. However we cannot argue that this applies to all Unix-like systems: The monolithic Linux kernel, Kerberos, httpd, SAMBA, X windowing, heck even OpenSSL. There’s many examples of tooling built on top of Unix systems that don’t follow that philosophy.

The traditional Unix way of doing things is definitely very outdated though.

Depends on what you mean. “Everything is a file”? Sure, that metaphor can be put to rest. “Low coupling, high cohesion”? That’s even more valid now for cloud architectures. You cannot scale a monolith efficiently these days.

In the end, Kubernetes is trying to impose a semi-distributed model of computation on a very NOT distributed operating system to the detriment of system complexity, maintainability, and security.

Kubernetes is more complex than a single Unix system. It is less complex than manually configuring multiple systems to give the same benefits of Kubernetes in terms of automatic reconciliation, failure recovery, and declarative configuration. This is because those three are first class citizens in Kubernetes, whereas they’re just afterthoughts in traditional systems. This also makes Kubernetes much more maintainable and secure. Every workload is containerized, every workload has predeclared conditions under which it should run. If it drifts out of those parameters Kubernetes automatically corrects that (when it comes to reconciliation) and/or blocks the undesirable behaviour (security). And Kubernetes keeps an audit trail for its actions, something that again in Unix land is an optional feature.

If you work with the Kubernetes model then you spend 10% more time setting things up and 90% less time maintaining things.

9P is much simpler and more elegant than HTTP

It also has negligible adoption compared to HTTP. And unless it provides an order of magnitude advantage over HTTP, then it’s going to be unlikely that developers will use it. Consider git vs mercurial. Is the latter better than git? Almost certainly. Is it 10x better? No, and that’s why it finds it hard to gain traction against git.

A filesystem does not exclusively mean an on-disk representation of a tree of files with a single physical point of origin. A filesystem can be just as “highly available” and distributed as any other way of representing resources of a system if not more so because of its abstractness.

Even an online filesystem does not guarantee high availability. If I want highly available data I still need to have replication, leader election, load balancing, failure detection, traffic routing, and geographic distribution. You don’t do those in the filesystem layer, you do them in the application layer.

Also, you’re “disappointed” in me? Lmao

Nice ad hominem. I guess it’s rules for thee, but not for me.

And how do you manage containers? With bespoke tools and infrastructure removed from the file abstraction. Which is another way Kubernetes is removed from the Unix way of doing things. Unless I’m mistaken, it’s been a long time since I touched Kubernetes.

So what’s the problem? Didn’t you just say that the Unix way of doing things is outdated? Let the CSI plugin handle the filesystem side if things, and let Kubernetes focus on the workload scheduling and reconciliation.

It’s not a preconception. They engaged with your way of doing things and didn’t like it.

Dismissal based on flawed anecdote is preconception.

By what standard? The standard of you and your employer? In general, you seem to be under the impression that the conventional hegemonic corporate “cloud” way of doing things is the only correct way and that everyone else is unskilled and not flexible.

No. I’m not married to the “cloud” way of doing things. But if someone comes to me and says “Hey boblin, we want to implement something on system foo, can you help us?” and I am not used to doing things the foo way I will say “I’m not familiar with it but let’s talk about your requirements, and why you chose foo” instead of “foo is for bureaucrats, I don’t want to use it”. I’d rather hire an open-mined junior than a gray-bearded Unix wizard that dismisses anything unfamilar. And I will also be the first person to reject use cases for Kubernetes when they do not make sense.

just that you should be more open-minded and not judge everyone else seeking a different path to the conventional model of cloud/distributed computing as naive, unskilled people making “bad-faith arguments”.

There are scenarios where cloud compute just does not make sense, like HPC. If the author had led with something like that, then they would have made a better argument. But instead they went for

cloud-native tooling feels like it’s meant for bureaucrats in well-paid jobs,

,

In the 90s my school taught us files and folders when we were 8 years old

, and

When you finally specify all those flags, neatly namespaced with . to make it feel all so very organised, you feel like you’ve achieved something. Sunk-cost fallacy kicks in: look at all those flags that I’ve tuned just so - it must be robust and performant!

It’s hard to not take that as bad faith.

I probably did go a bit ad hominem in my last paragraph. By the time I was done with the article I was very frustrated by what seemed to be some very bad faith arguments (straw man, false dilemma) that were presented.

This vmalert tool is just an interface to another, even more complicated piece of software.

Not really just an interface. It is a pluggable service that connects to one or more TSDBs, performs periodic queries, and notifies another service when certain thresholds are exceeded. So with all those configuration options, why is the standalone binary expected to have defaults that may sound same on one system but insane in a different one? If the author wants out of the box configuration they could have gotten the helm chart or the operator and then that would be taken care of. But they seem to be deathly allergic to yaml, so I guess that won’t happen.

Since when do Unix tools output 3,000 word long usage info? Even GNU tools don’t even come close…

You just said that this software was much more complex than Unix tools. Also if only there were alternate documentation formats….

HTTP and REST are very strange ways to accomplish IPC or networked communication on Unix when someone would normally accomplish the same thing with signals, POSIX IPC, a simpler protocol over TCP with BSD sockets, or any other thing already in the base system.

Until you need authentication, out of the box libraries, observability instrumentation, interoperability… which can be done much more easily with a mature communication protocol like HTTP. And for those chasing the bleeding edge there’s gRPC.

I would hope the filesystems you use are “high availability” lol

They’re not, and I’m disappointed that you think they are. Any individual filesystem is a single point of failure. High availability lets me take down an entire system with zero service disruption because there’s redundancy, load balancing, disaster recovery…

the humble file metaphor can still represent these concepts

They can, and they still do… Inside the container.

It’s not a lack of skill as your comment implies but rather a rejection of this way of doing things.

Which I understand, I honestly do. I rejected containers for a (relatively) long time myself, and the argument that the author is making echoes what I would have said about containers. Which is why I believe myself to be justified in making the argument that I did, because rejecting a way of doing things based on preconception is a lack of flexibility, and in cloud ecosystems that translates to a lack of skill.

Literally copied and pasted that from the article.

I am someone with kubernetes in my job title. If you as a developer are expected to know about kubernetes beyond containerizing your application then your company has set itself up for failure. As you aptly said kubernetes is an ecosystem, and the dev portion is a small niche of that.

You can’t run vmalert without flags

Running grep without parameters is also pretty fucking useless.

500 words in to the over 3,000 word dump, I gave up.

Claims to have a Unix background, doesn’t RTFM.

Nobody really uses Kubernetes for day-to-day work, and it shows. Where UNIX concepts like files and pipes exist from OS internals up to interaction by actual people, cloud-native tooling feels like it’s meant for bureaucrats in well-paid jobs.

Translation: Author does not understand APIs.

Want an asynchronous, hierarchical, recursive, key-value database? With metadata like modified times and access control built-in? Sounds pretty fancy! Files and directories.

Ok. Now give me high availability, atomic writes to sets of keys, caching, access control…

I’m ashamed enough that I can’t really apply to these jobs

This reads as “I applied to the jobs and got rejected. There’s nothing wrong with me, so the jobs must be broken”.

Wouldn’t he consider the collective the upper class, and any non-Borg the commoners? Although he might be at odds with the whole “assimilate” thing.